Abstract
What is the physiological basis of longterm memory? The prevailing view in Neuroscience attributes changes in synaptic efficacy to memory acquisition, implying that stable memories correspond to stable connectivity patterns. However, an increasing body of experimental evidence points to significant, activityindependent fluctuations in synaptic strengths. How memories can survive these fluctuations and the accompanying stabilizing homeostatic mechanisms is a fundamental open question. Here we explore the possibility of memory storage within a global component of network connectivity, while individual connections fluctuate. We find that homeostatic stabilization of fluctuations differentially affects different aspects of network connectivity. Specifically, memories stored as timevarying attractors of neural dynamics are more resilient to erosion than fixedpoints. Such dynamic attractors can be learned by biologically plausible learningrules and support associative retrieval. Our results suggest a link between the properties of learningrules and those of networklevel memory representations, and point at experimentally measurable signatures.
Introduction
The ability to form and retain memories of past experience is fundamental to behavior, supporting adaptable responses, and future planning^{1}. These internal representations persist over extended durations and may be reactivated by appropriate retrieval cues^{2}. Currently, it is widely accepted that synaptic connections between neurons play a central role in the physiological basis of longterm memory storage^{3} (see refs. ^{4,5} for other possibilities). The process of learning, on its part, is understood as stimulusdriven neural activity sculpting network architecture, i.e., Hebbian plasticity^{6}.
If an internal memoryrepresentation is stable over time, then one could assume that some properties of its underlying neural implementations also exhibit invariance over this period. However, at the level of single synapses, no such robustness exists (reviewed in refs. ^{7,8,9,10}). Over the past decade, several studies, both ex vivo^{11,12} and in vivo^{13}, suggested that synapses undergo significant spontaneous changes. These fluctuations persist even in the absence of neural activity, with magnitude estimated to be as large as that of directed, Hebbian, plasticity^{14}.
How, then, can memory traces remain stable over time? What remains invariant while all synaptic strengths vary? Various studies have proposed candidate invariant features, at different levels of organization of neural networks. For single synapses, invariance may be implemented in a subset of the largest spines^{13,15}. Invariance may, instead, only emerge at the level of the connection between neurons, typically comprising several synapses. This allows individual synapses to fluctuate, under the constraint of stable overall connection strength between two cells^{16,17}. Higher up the organizational hierarchy, invariant features may manifest only at the network level. This would allow individual connections to fluctuate, while at the same time some network properties remain stable^{9}.
In this work, we suggest that such networklevel invariance arises naturally when considering the effect of homeostatic mechanisms on network stability. Stability is best understood by considering a global property of the network connectivity matrix—its eigenvalues. This is a collection of points in the complex plane, each having a real and imaginary part. We show that under activityindependent synaptic fluctuations and known homeostatic mechanisms, information encoded in the real part of these eigenvalues is strongly eroded, while imaginarycoded information is spared. Consequently, in a system that utilizes imaginarycoded memory, single synapses may exhibit ongoing fluctuations whereas global features remain invariant as networklevel properties.
We investigate this concept by showing how different homeostatic plasticity mechanisms degrade real or imaginarycoded memories, and how spike timing dependent plasticity (STDP) can store transient inputs as imaginarycoded memories. We then show the implications of such memories—that learned representations give rise to stable oscillatory trajectories of network activity. These memory states can be viewed as the timevarying analogs of stable fixed points in the classic Hopfield model^{18}. Having been learned and embedded in a component of connectivity, memory items may be transiently retrieved by supplying an associative recall cue.
Our results suggest a principle by which memory can be learned and retained in a stable manner despite significant ongoing synaptic fluctuations. The implications of such a mechanism to experimental data are discussed both in terms of measured neural activity and in terms of synaptic plasticity during learning as opposed to at rest.
Results
Model
Our model is based on a standard framework of firingrate neural networks^{19}. It consists of N recurrently connected neurons, with W_{ij} the synaptic connection strength from neuron j to i. Each neuron i transforms its input x_{i} into firing rate via a nonlinearity ϕ(x_{i}), where the state vector x = (x_{1} x_{2} ⋯ x_{N})^{⊤} evolves as
and b is an external input. Here and below we denote by ϕ(x) the vector obtained by applying ϕ to each coordinate of x.
Connectivity of taskperforming networks is often designed to achieve a desired functionality, and assumed to be constant while the network is performing the task^{18,20}. There are models in which connectivity coevolves with neural dynamics, but changes are usually confined to a training phase, whereas connectivity is kept constant during the test phase^{21,22}. These models are consistent with the expectation of synaptic tenacity in the absence of learning. In our model, to incorporate the recent observations on synaptic fluctuations, the connectivity matrix W continuously coevolves with neural activity x throughout all task phases, albeit with a slower timescale (Fig. 1, see also ref. ^{23}).
In order to study the coexistence of memory with synaptic fluctuations, we let W evolve due to contributions arising from both learningrelated and fluctuationrelated terms, denoted by Δ_{L} and Δ_{F} respectively:
with η > 0 the plasticity rate (relative to neural dynamics). The fluctuation term includes stochastic, activityindependent noise in synaptic strength, as well as a homeostatic mechanism to control synaptic and firingrate stability. These are precisely the processes which endanger the resilience of an acquired memory that is assumed to be stored in synaptic patterns. We first consider how an existing memory is eroded by these processes, and later address the learning part and the interplay between the two.
Homeostatic plasticity erodes realcoded information
We model spontaneous activityindependent synaptic fluctuations by a white noise process ξ_{ij} driving each synapse ij independently. The variance of such a process grows without bound, and thus, without a restraining mechanism these fluctuations would lead to divergence of the synaptic weights W_{ij}. A plausible restraining agent is homeostatic plasticity, modeled here as a networklevel mechanism that stabilizes both synaptic weights and neural activities on average over long timescales. To understand this stabilization, we turn to dynamical systems theory that determines stability about a setpoint by the spectrum of the appropriate Jacobian matrix (which is the local linear approximation of the dynamics). This spectrum consists of a set of complex eigenvalues—a collection of points in the complex plane, each having a real part and an imaginary part.
In general, the real part of this spectrum defines the system’s stability: a system is only stable if all its eigenvalues have negative real parts. The imaginary part of the spectrum, in contrast, determines the typical timescales of smallamplitude dynamics around this setpoint, but not stability itself (Fig. 2a). Therefore, while the real part of the spectrum must be under the control of homeostatic plasticity, its imaginary part is not constrained by the requirement of stability, and is free to store information (Fig. 2b).
The arguments above derive from a general intuition on system stability; they are not a mathematical proof, as they depend on the existence of a setpoint and its exact properties. They do, however, provide motivation to test this idea using various homeostatic mechanisms. We perform such tests using the connectivity matrix W as a proxy for the Jacobian. In the case of a linear network, or of linearizing around the origin, the two are equivalent. Our results below indicate that such an approximation is useful also in more general cases.
The embedding of a memory item is often represented in learning theory as a lowrank perturbation to the connectivity matrix, W → W + δW. For example, in the Hopfield model a memory is associated with a particular pattern of activity, a vector u ∈ ℝ^{N}, and is embedded in connectivity by adding a perturbation of the form δW = uu^{⊤}. Such a perturbation is symmetric, δW_{ij} = δW_{ji}, and modifies the connectivity spectrum to include a real positive eigenvalue (with an eigenvector in the direction of u). To tap into the resilience of the imaginary part of the spectrum as argued above, one might add an antisymmetric perturbation (δW_{ij} = −δW_{ji}) that gives rise to an imaginary conjugate pair of eigenvalues. This defines a different type of memory item, the simplest form being δW = uv^{⊤} − vu^{⊤}. If the above general arguments on system stability are correct, such memory items should be more resilient to synaptic fluctuations. We test this by comparing the erosion of the two types of memory items under various homeostatic mechanisms. We first embed memories corresponding to either real or imaginary eigenvalues into the connectivity matrix W, and then follow the dynamics of Eqs. (1) and (2) without active learning (Δ_{L} = 0), but with various homeostatic models in Δ_{F}.
Perhaps the simplest implementation of a homeostatic mechanism is by dissipative synaptic dynamics. Together with the noise ξ, this gives a fluctuation term
with β > 0 the rate of dissipation. Figure 2c shows the eigenvalues of the connectivity matrix as a function of time (gray lines), with the eigenvalues corresponding to the memory highlighted in green. It is seen that the memory representation rapidly decays for both real (top) and imaginary (bottom) eigenvalues. This is expected from a dissipative system, where all information decays exponentially with a rate β. Therefore, in the presence of such a mechanism, neither type of memory items can be sustained for longer than the decay time 1/β. However, as will be shown below, this is not the case for more indirect homeostasis mechanisms.
A biologically plausible homeostasis mechanism can be modeled as an activitydependent rule—where the synaptic matrix is modified to achieve a stable postsynaptic firingrate^{24,25,26}:
with ϕ_{0} an arbitrary targetrate vector, and \(\circ\) denoting a Hadamard (elementwise) product. Stabilization of firing rates around the setpoint ϕ_{0} is achieved in this case by scaling the weight of each synapse by a factor dependent on its pre and postsynaptic neurons. Although not explicitly symmetric (which would lead to a change in real eigenvalues), the stabilization induced by this rule requires control over the real part of the relevant Jacobian. Accordingly, Fig. 2d (top) shows that memories stored as real eigenvalues of W rapidly decay. Imaginarycoded memories, on the other hand, may persist indefinitely without interfering with homeostasis (Fig. 2d, bottom).
Finally, inspired by ref. ^{27}, we consider a homeostasis mechanism that does not have a welldefined firingrate setpoint. Instead, this rule contains an antiHebbian term that reduces the connections between correlated neurons, thus pushing towards decorrelated firing rates across the network:
where ϕ_{pre}, ϕ_{post} are two sigmoidal functions and I is the identity matrix. This rule leads to persistent and constrained fluctuations in both connectivity and firing rates^{27}. The activity x is dominated by the unstable modes of W, which are then suppressed by the antiHebbian term, leading to a new unstable mode in an endless succession.
Once again, we find that the decay of imaginarycoded memories is orders of magnitude slower than that of realcoded ones (Fig. 2e). Note that if the sigmoidal functions are identical, ϕ_{pre} = ϕ_{post}, this rule can only modify the symmetric part of W. In practice, for many nonidentical choices of the sigmoid functions, the modification is still mostly symmetric. Nevertheless, the relative decay of imaginary and real based memories is similar to the case of the ratecontrol rule, that does not have any symmetric tendency.
Some evidence for the generality and limits of validity of these results is presented in Supplementary Fig. 1, where sparse networks are considered. The amplitude of embedded imaginary memories decreases smoothly as the network becomes sparser; they remain more resilient than realcoded memories for the entire regime tested.
In light of these results, a natural question arises: can a dynamical learning rule utilize the imaginary subspace to robustly code and store memory representations? Perhaps surprisingly, we find that Spike Time Dependent Plasticity (STDP), a well documented and biophysically plausible learning rule, provides a natural candidate for learning imaginarycoded memories.
STDP stores imaginarycoded information
Symmetric and antisymmetric matrices give rise to real and imaginary eigenvalues respectively. It is thus reasonable that an antisymmetric modification to the synaptic weight matrix W would primarily lead to changes in the imaginary part of its spectrum. Local learning rules observed in experiments (e.g., STDP) have a welldefined directionality: consecutive firing of neuron j before i leads to a strengthening of the connection W_{ij} and to the weakening of the reverse connection. The temporal asymmetry of STDP^{28} leads to an approximately antisymmetric learning rule when applied to our rate model (see “Methods”); as such, this rule mostly affects the imaginary part of the spectrum. In the case of perfect antisymmetry, we find the form Δ_{L} = ϕy^{⊤} − yϕ^{⊤}, which modifies only the antisymmetric component of W. The vector y is a smoothed version of the activity ϕ(x), and arises in our ratebased formulation when applying an exponential STDP kernel to spiketrains (see “Methods”).
These arguments suggest that a biologically motivated learning rule naturally stores imaginarycoded information, thereby rendering it relatively resilient to the effect of homeostatically controlled synaptic fluctuations. We will next investigate how such a memory can be acquired, retained and retrieved in the presence of synaptic fluctuations. For simplicity, we will use the purely antisymmetric Δ_{L}.
The acquisition, i.e., the encoding and storage of a new memory trace, is initiated by stimulating the network with an external signal, b(t). A matrix with imaginary eigenvalues is necessarily of (at least) rank 2, corresponding to a twodimensional space spanning the memory representation. We therefore present the network with a randomly timevarying input evolving on a plane spanned by two arbitrary directions u, v ∈ ℝ^{N} (see “Methods”). As the input drives neural activity x onto the (u, v) plane, the activitydependent learning operator Δ_{L} follows and becomes nonnegligible, which in turn causes a change in connectivity.
The learning procedure stores geometric information of the external stimulus, specifically the directions u and v, within the antisymmetric part of the connectivity matrix. This encoding is manifested as a rank2 operator uv^{⊤} − vu^{⊤} which is embedded into W. To see this, we follow the spectrum of W as a function of time: during stimulus presentation, a complex conjugate eigenvalue pair forms (Fig. 3a), with corresponding eigenvectors overlapping completely with the plane spanned by u, v (Fig. 3b). The strength of the memory representation—corresponding to the magnitude of the imaginary eigenvalue—depends monotonically on stimulation duration and on input amplitude. At later times, additional stimuli may be stored using the same learning protocol (Fig. 3c).
The nature of imaginarycoded memories
We have seen that a biologically plausible learning rule can capture the orientation in neuronal statespace of an incoming stimulus, and encode this information as a pair of imaginary eigenvalues in the network connectivity matrix. What is the nature of this memory in terms of network activity? We find that learning creates attractors in statespace, similar in fashion to those in the Hopfield model^{18}. However, rather than fixed points, here the attractors are timevarying stable states—namely, limit cycles. To see this most clearly, we consider a single imaginarycoded memory embedded in the network, and examine neural dynamics while keeping W fixed. Following the Hopfield paradigm, we write W as:
where the coefficient ρ > 0 represents the strength of the memory representation^{29}.
With one stored memory as in Eq. (6), we find that, from any nonzero initial condition, the dynamics converge to periodic motion concentrated on the ‘memory plane’ spanned by u and v. Figure 4a depicts the projections of neural activity on this plane, for two initial conditions (light gray trajectories), both converging to the limitcycle attractor (dark closed trajectory). An approximate lowdimensional description of this limit cycle can be obtained for an infinitely steep nonlinearity ϕ (i.e., a stepfunction). The full dynamics are then well approximated by their projected coordinates on the plane, p_{u} and p_{v}:
where \(q_{\mathbf{v}} \approx {\mathrm{arctan}}\left( {\frac{{p_{\mathbf{v}}}}{{p_{\mathbf{u}}}}} \right)\). This lowdimensional system exhibits a stable limitcycle around the origin (see Supplementary Note 1 and Supplementary Fig. 3). We conclude that imaginarystored memory items correspond to dynamic attractors, with geometry defined by that of the stimulating input. This behavior stands in contrast to the classic—symmetric—Hopfield model, where memories are represented by fixedpoint attractors corresponding to fixed values of neural activity.
Embedding multiple memory planes \(\{ {\mathbf{u}}^{\left( k \right)},{\mathbf{v}}^{\left( k \right)}\} _{k = 1}^M\) corresponds to setting
where the columns of U are the memory patterns (interleaved u^{(k)} and v^{(k)}), and D is a 2M × 2M blockdiagonal matrix, with the kth block reading \(\left( {\begin{array}{*{20}{c}} 0 & {\rho _k} \\ {  \rho _k} & 0 \end{array}} \right)\). Now, a locally stable limitcycle lies on each embedded plane, and the network functions as an associative memory: initiating the dynamics within the basin of attraction of one plane—providing the network with partial information of the memory to be retrieved—leads to the recovery of the full memory item (Fig. 4b). Similar to the Hopfield model, the memory capacity is found to be proportional to system size^{30}. Numerical simulations presented in Supplementary Fig. 4 show that in fact the proportionality constant is slightly higher compared with that of the symmetric Hopfield model (when normalized by a factor of two, since each memory resides on a plane; see Supplementary Note 2).
Life cycle of a memory trace
We next consider the entire lifecycle of a memory in the presence of synaptic fluctuations and homeostasis, starting from learning, through retention and to retrieval. During a learning event, implemented by presenting a stimulus in the twodimensional memory plane, a memory representation is formed by the Hebbian learning rule. Figure 5a (left) shows the overlaps of neural activity onto the two planes, r_{1} (green) and r_{2} (blue), together with the stimulus which drives learning (shades). These projections are elevated during stimulation, which—via the Hebbian learning rule—modifies the synaptic matrix to store each of the planes in connectivity. The projection r_{3} onto a third plane, which was not learned, is negligible as shown in the bottom line (orange).
After learning, the two memory items are stored as pairs of imaginary eigenvalues, remaining stable over time, until they are retrieved at times t_{1} and t_{2}, respectively. At retrieval, activity is transiently attracted to the respective memory planes, as indicated by the spikes in the overlaps (Fig. 5a, right). During retrieval, activity follows the stored dynamic trajectory, exhibiting its typical oscillations (Fig. 5a, right, blue zoom). At the same time, the projection onto an arbitrary plane shows no temporal structure (orange zoom). Finally, stimulating the network with a novel cue does not elicit a significant response in neural activity in any of the projections (t_{3} in Fig. 5a).
The effect of retrieval on the connectivity, namely on the stored memory itself, is somewhat unpredictable and depends on the exact state of the network and on the memory properties. As an example, in Fig. 5c it is seen that the green memory is damaged by retrieval, namely the magnitude of the corresponding imaginary eigenvalue is decreased. This may be caused by the homeostatic mechanism that constrains activity, in particular the component projected onto the memory plane by the retrieval event. In contrast, the blue memory is slightly strengthened by retrieval, as seen by the increased magnitude of the eigenvalue pair. In other cases the memory remains unaffected.
Throughout this entire cycle, synapses fluctuate under the effect of activityindependent noise and homeostasis. Figure 5d shows a few example synapses tracked across time, during both phases. We may disentangle the two effects, spontaneous and activitydependent, and estimate their relative contribution to synaptic fluctuations; their magnitude is found to be similar (Fig. 5b), in agreement with the experimentally observed phenomenon^{14}.
Discussion
Experimental evidence on the perpetual changes in synaptic strengths, with and without relation to activity or learning, has accumulated by now to form a well accepted picture: synapses are not as stable as once thought. Earlier theoretic work studied the statistical properties of single synapse fluctuation using phenomenological models^{12,31,32}. These models were successful in capturing the quantitative statistics of single synapses along time and across a population, but did not address their context within an active network. The more difficult question of the implications of such fluctuations to network functionality has been highlighted in several recent reviews^{7,8,9,10}. Neural network models that can store memory provide a framework for addressing this problem. This can be done by adding noise and investigating its effect on memories and network dynamics.
Previous theoretical work has established the robustness of the classic Hopfield model to static synaptic noise^{33,34}, which exhibits a smooth and moderate decline of the critical memory load as a function of the noise variance. However, introducing a dynamic noise source to the synaptic weight evolution has a qualitatively different effect. In this case, the noise typically causes the weights to diverge with time, thus erasing any stored memory items. As shown in Fig. 2c, a dissipative term can stabilize the stochastic fluctuations, but will also erase any memory in the network. Here, we showed that it is possible to store, retain and retrieve memories in a recurrent neural network despite significant ongoing synaptic fluctuations. Motivated from a fundamental systemstheory perspective, we argued that fluctuations, and homeostasis mechanisms that control them, place a strong constraint only on the real part of the eigenvalues of the connectivity matrix. A corollary of this observation is that memories constituting stable dynamical trajectories of neural activity, associated with imaginary eigenvalues, can be kept encoded in connectivity for extended times while synapses fluctuate under noise and homeostasis.
We implemented this idea with the simplest form of such memory: a twodimensional plane in which periodic activity persists as a stable limit cycle. This implementation extends the Hopfield picture, where memories are fixedpoint attractors with static activity patterns, to memories represented by dynamic trajectories in neural state space^{29}. This temporal dependence is more consistent with experimental data: the oscillatory trajectories of memorytrace activations that arise naturally in our model resemble networklevel oscillations observed during memory retrieval^{35} and consolidation^{36}.
Since imaginary eigenvalues are associated with the antisymmetric component of connectivity, the observed asymmetry of STDP^{6} naturally suggests that memory items of this type can be learned dynamically. We have demonstrated how such learning occurs by a single stimulus presentation which spans a twodimensional plane in activity space, thus allowing the embedding of periodic motion. This is most simply shown for a perfectly antisymmetric STDP kernel, but is valid as long as the kernel contains a significant antisymmetric component. The effects of symmetry of STDP on memory retention have also been noted in a different modeling context. In^{23} the Hopfield model was studied in the presence of ongoing STDP; it was found that unstructured noise inserted into the neural state could stabilize memories with antisymmetric, but not with symmetric, learning.
From a more general perspective, any learning rule represents the interaction of the system with its environment; if this rule is not homogeneous in space and time, the signature of this interaction might be encoded in some subspace of network connectivity—a component with particular symmetries. This would allow an invariant subspace of connectivity within which memories are stored, and which is minimally tampered by homeostatic fluctuations. Indications of such invariant features have been recently observed experimentally^{37}: while individual neurons exhibited significant change in their activity patterns relative to behavior in a decision task, population activity and behavior remain stable over weeks. The notion of an invariant subspace has also been suggested to underlie stable behavior during working memory tasks, despite ongoing neural activity^{38}.
It is also possible that other principles can be formulated that allow coexistence of stable function with synaptic fluctuations. For example, the microscopic degeneracy of representation was shown to support stable input–output relations in a feedforward network amid strong fluctuations^{39}. Although the model and its implementation are very different from ours, the motivating question is broadly similar. More recently it was proposed that in balanced cortical networks, inhibitory connectivity alone bears the burden of robust information storage^{40}, thereby rendering memories insensitive to fluctuations of excitatory synapses.
Our model was based on general considerations of system stability, without relying on specific implementations of homeostasis. Nevertheless, it gives rise to two points worthy of discussion in relation to experimental predictions. First, imaginarycoded memories give rise to limitcycle attractors. Thus, retrieval of an item from longterm memory to working memory should give rise to oscillatory activity. These signatures of oscillations might be detected from the spectral properties of neural activity, expected to vary significantly between learning and rest phases (see Supplementary Note 3 and Supplementary Fig. 5). Such signatures have already been observed^{41,42} but our model suggests an additional feature that may be hiding in the data. The magnitude of the imaginary eigenvalue should correlate with both the oscillation frequency and with memory strength. Indeed, for a given memory item in our model, a spectral analysis correlates almost perfectly with memory strength. Considering many different memory items, we find that, despite interitem variability of this slope, the two measurable quantities maintain a high correlation (see Supplementary Note 3 and Supplementary Fig. 5).
Our model could, in principle, be tested by directly examining synaptic strengths during learning. The model predicts that learningrelated plasticity should be preferentially antisymmetric. This could be checked by monitoring the synaptic strengths between reciprocally connected neurons during learning and rest, and our model predicts measurable differences in these phases. The exact correlation structure, however, is not expected to be universal and could reflect details of the homeostatic mechanisms (see Supplementary Note 3 and Supplementary Fig. 6).
More generally, our results suggest that much systemslevel understanding can be gleaned by measuring and analyzing a population of synaptic strengths across time in large networks. Specifically, beyond the statistical analysis of the single synapse, invariant structure in the highdimensional space of connectivity should be searched. Moving towards such an understanding will hopefully be possible with the advancement of experimental techniques, that will allow monitoring of multiple synapses across extended times and during various phases of behavior.
Methods
Model simulation
We use homemade MATLAB software in order to numerically simulate Eqs. (1) and (2). The spectra of matrices are computed using builtin MATLAB functions, and their timeseries sorted using the eigenshuffle.m MATLAB script by John D’Errico, freely available online. For the nonlinearity in firing rates, we use the hyperbolic tangent function, ϕ(z) = tanh(z). We have verified that the results presented in Fig. 2 are reproduced also with a rectifiedlinear input–output function, ϕ(z) = max(−5, z), where a negative threshold is required for achieving homeostasis. For all figures we simulate a network with N = 128 neurons, unless otherwise specified. Simulations with larger networks similarly exhibit all of the discussed phenomena. For numerical integration we use a time constant dt = 0.1. Synaptic weights were evolved with a plasticity rate η = 0.01, this includes learning and homeostatic plasticity rules. For the lowpass filter y, we use a firstorder filter with timescale τ = 50 (see next “Methods” section). For the learning process we use the timedependent input b(t) = c_{u}(t)u + c_{v}(t)v with \(u_{i},v_{i}\sim {\cal{N}}\left(0,\frac{1}{N} \right)\) and independent. The timedependent functions c_{u}(t) and c_{v}(t) each follow an independent OrnsteinUhlenbeck process with timescale 0.01. For retrieval, a brief (2 simulation time constants) pulse in the direction of u is applied to the network, namely c_{u}(t) = 10 and c_{v}(t) = 0.
For the ratecontrol homeostasis rule, we draw each component of the targetrate vector ϕ_{0} independently from a uniform distribution over the interval [−1, 1]. For the decorrelation homeostasis rule, we use ϕ_{pre}(x) = ϕ(x) = tanh(x) and \(\phi _{{\mathrm{post}}}\left( {\mathbf{x}} \right) = {\mathrm{tanh}}\left({\mathbf{x}}  \overline {\mathbf{x}} \right)\), where \(\overline {\mathbf{x}}\) is a firstorder lowpassed version of x, with timescale τ_{x} = 20. In all cases, we model spontaneous fluctuations by a white noise process, \(\xi _{ij}\left(t\right)\sim {\cal{N}}\left(0,\frac{1}{N}\right)\), independent across time and synapses ij. For the dissipative synaptic dynamics we use β = 0.1.
In Fig. 3b, we compute the overlap between two planes, spanned by u_{1}, v_{1} and u_{2}, v_{2} respectively, as follows: Compute the projection of each spanning vector of the first plane onto the second plane:
and then the overlap is given by
In Fig. 4, we slightly modified the imaginarycoded memory representation Eq. (6). In particular, we set connectivity to W = ρ(uv^{⊤} − vu^{⊤}) + γ(uu^{⊤} + vv^{⊤}), with γ > 1. The second term emerges naturally in W when the memory is learned via our dynamic learning protocol; without it, the origin in phasespace ℝ^{N} becomes a locally stable fixed point and trajectories decay. Numerically, in the limit of small integration step dt → 0, we find that, for γ = 0, the origin is actually globally stable, and the memoryrelated limitcycle disappears. On the other hand for discretetime dynamics, the model with γ = 0 is stable, and this is the version used for the capacity calculations.
For generating Fig. 5b we compute the contribution of each plasticity term Δ(t) as the temporal average of \(\frac{1}{{N^2}}\mathop {\sum}\limits_{ij} {\Delta _{ij}(t)}\), from a simulation of our model with the decorrelation homeostasis rule.
Derivation of the Hebbian learning rule
In this section we derive the ratebased learning rule Δ_{L}. Our starting point is a Poisson spiking neuron with output spiking activity given by a train of point events \(S_i\left( t \right) = {\sum} \delta (t  t_k)\) that follow from a timedependent firing rate ϕ_{i}(t)^{43}, with δ(t) the Dirac delta function. STDP learning is characterized by a differential update of the synaptic efficacy W_{ij}, based on the temporal distance Δt between spiking of unit i and unit j; the amplitude of change is given by the ‘learning window’, or kernel, K(Δt)^{28}. In our rate model we are interested in variations slow compared to interspike intervals, and therefore we average the effect of STDP over this timescale. The average quantity which describes the temporal relation between inbound and outbound neural activity is the correlation function,
where angular brackets denote ensemble averaging over the spiking activity and overbar denotes temporal averaging. STDP learning can then be formalized as the convolution of this correlation function with the kernel^{23}:
Within our slowtimescale approximation, correlations arise from local changes in firing rate rather than differences between individual spikes. Therefore, in terms of the instantaneous firing rates, C_{ij}(t;s) ≈ ϕ_{i}(t)ϕ_{j}(s)^{43}. Using a learning kernel of the form
where τ_{P}, τ_{D} are the two timescales of the kernel, and a_{P} > 0 and a_{D} < 0 are the positive and negative amplitude for increase or decrease of the connectivity, depending on the sign of correlation, we carry out the integration in Eq. (9) to obtain
where y^{P}, y^{D} are firstorder lowpass filters of spiking rates ϕ, with respective timescales τ_{P} and τ_{D}.
In general, the parameters of K give rise to an asymmetric learning operator Δ_{L}. The extent of asymmetry is determined by the discrepancy between the two pairs of kernel parameters, i.e. the difference in timescales of potentiation and depression τ_{P}, τ_{D}, and the two amplitudes a^{P}, a^{D}. When τ_{D} = τ_{P} and a_{D} = −a_{P}, the learning operator is purely antisymmetric:
Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Code availability
Example code for simulating our main results can be found at https://github.com/lsusman/stablememory.
Data availability
Data sharing not applicable to this article as no datasets were generated or analysed during the current study.
References
Clayton, N. S., Bussey, T. J. & Dickinson, A. Can animals recall the past and plan for the future? Nat. Rev. Neurosci. 4, 685 (2003).
Kahana, M. J. Foundations of Human Memory. (OUP, USA, 2012).
Poo, M.M. et al. What is memory? the present state of the engram. BMC Biol. 14, 40 (2016).
Tsien, R. Y. Very longterm memories may be stored in the pattern of holes in the perineuronal net. Proc. Natl Acad. Sci. USA 110, 12456–12461 (2013).
Titley, H. K., Brunel, N. & Hansel, C. Toward a neurocentric view of learning. Neuron 95, 19–32 (2017).
Bi, G.Q. & Poo, M.M. Synaptic modification by correlated activity: Hebb’s postulate revisited. Annu. Rev. Neurosci. 24, 139–166 (2001).
Mongillo, G., Rumpel, S. & Loewenstein, Y. Intrinsic volatility of synaptic connections–a challenge to the synaptic trace theory of memory. Curr. Opin. Neurol. 46, 7–13 (2017).
Clopath, C., Bonhoeffer, T., Hübener, M. & Rose, T. Variance and invariance of neuronal longterm representations. Philos. Trans. R. Soc. B 372, 20160161 (2017).
Chambers, A. R. & Rumpel, S. A stable brain from unstable components: emerging concepts and implications for neural computation. Neuroscience 357, 172–184 (2017).
Ziv, N. E. & Brenner, N. Synaptic tenacity or lack thereof: spontaneous remodeling of synapses. Trends Neurosci. 41, 89–99 (2018).
Minerbi, A. et al. Longterm relationships between synaptic tenacity, synaptic remodeling, and network activity. PLoS Biol. 7, e1000136 (2009).
Yasumatsu, N., Matsuzaki, M., Miyazaki, T., Noguchi, J. & Kasai, H. Principles of longterm dynamics of dendritic spines. J. Neurosci. 28, 13592–13608 (2008).
Loewenstein, Y., Yanover, U. & Rumpel, S. Predicting the dynamics of network connectivity in the neocortex. J. Neurosci. 35, 12535–12544 (2015).
Dvorkin, R. & Ziv, N. E. Relative contributions of specific activity histories and spontaneous processes to size remodeling of glutamatergic synapses. PLoS Biol. 14, e1002572 (2016).
Yang, G., Pan, F. & Gan, W.B. Stably maintained dendritic spines are associated with lifelong memories. Nature 462, 920 (2009).
Koester, H. J. & Johnston, D. Target celldependent normalization of transmitter release at neocortical synapses. Science 308, 863–866 (2005).
Fares, T. & Stepanyants, A. Cooperative synapse formation in the neocortex. Proc. Natl Acad. Sci. USA 106, 16463–16468 (2009).
Hopfield, J. J. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl Acad. Sci. USA 79, 2554–2558 (1982).
Dayan, P. & Abbott, L. F. Theoretical Neuroscience: Computational And Mathematical Modeling of Neural Systems. (MIT Press, 2005).
BenYishai, R., BarOr, R. L. & Sompolinsky, H. Theory of orientation tuning in visual cortex. Proc. Natl Acad. Sci. USA 92, 3844–3848 (1995).
Sussillo, D. & Abbott, L. F. Generating coherent patterns of activity from chaotic neural networks. Neuron 63, 544–557 (2009).
Kurikawa, T. & Kaneko, K. Embedding responses in spontaneous neural activity shaped through sequential learning. PLoS Comput. Biol. 9, e1002943 (2013).
Wei, Y. & Koulakov, A. A. Longterm memory stabilized by noiseinduced rehearsal. J. Neurosci. 34, 15804–15815 (2014).
Buonomano, D. V. A learning rule for the emergence of stable dynamics and timing in recurrent networks. J. Neurophysiol. 94, 2275–2283 (2005).
Turrigiano, G. G. The selftuning neuron: synaptic scaling of excitatory synapses. Cell 135, 422–435 (2008).
Elboustani, S., Ip, J. P. K. & Bretonprovencher, V. Locally coordinated synaptic plasticity of visual cortex neurons in vivo. Science 1354, 1349–1354 (2018).
Magnasco, M. O., Piro, O. & Cecchi, G. A. Selftuned critical antihebbian networks. Phys. Rev. Lett. 102, 258102 (2009).
Bi, G.Q. & Poo, M.M. Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type. J. Neurosci. 18, 10464–10472 (1998).
Coolen, A. C. C. Theory of Neural Information Processing Systems. (Oxford University Press, 2005).
Amit, D. J., Gutfreund, H. & Sompolinsky, H. Storing infinite numbers of patterns in a spinglass model of neural networks. Phys. Rev. Lett. 55, 1530 (1985).
Loewenstein, Y., Kuras, A. & Rumpel, S. Multiplicative dynamics underlie the emergence of the lognormal distribution of spine sizes in the neocortex in vivo. J. Neurosci. 31, 9481–9488 (2011).
Statman, A., Kaufman, M., Minerbi, A., Ziv, N. E. & Brenner, N. Synaptic size dynamics as an effectively stochastic process. PLoS Comput. Biol. 10, e1003846 (2014).
Sompolinsky, H. Neural networks with nonlinear synapses and a static noise. Phys. Rev. A 34, 2571 (1986).
Tirozzi, B. & Tsodyks, M. Chaos in highly diluted neural networks. Europhys. Lett. 14, 727 (1991).
Skarda, C. A. & Freeman, W. J. How brains make chaos in order to make sense of the world. Behav. Brain Sci. 10, 161–173 (1987).
Paré, D., Collins, D. R. & Pelletier, J. G. Amygdala oscillations and the consolidation of emotional memories. Trends Cogn. Sci. 6, 306–314 (2002).
Driscoll, L. N., Pettit, N. L., Minderer, M., Chettih, S. N. & Harvey, C. D. Dynamic reorganization of neuronal activity patterns in parietal cortex. Cell 170, 986–999 (2017).
Druckmann, S. & Chklovskii, D. B. Neuronal circuits underlying persistent representations despite time varying activity. Curr. Biol. 22, 2095–2103 (2012).
Ajemian, R., D’Ausilio, A., Moorman, H. & Bizzi, E. A theory for how sensorimotor skills are learned and retained in noisy and nonstationary neural circuits. Proc. Natl Acad. Sci. USA 110, E5078–E5087 (2013).
Mongillo, G., Rumpel, S. & Loewenstein, Y. Inhibitory connectivity defines the realm of excitatory plasticity. Nat. Neurosci. 21, 1463 (2018).
Osipova, D. et al. Theta and gamma oscillations predict encoding and retrieval of declarative memory. J. Neurosci. 26, 7523–7531 (2006).
Raghavachari, S. et al. Gating of human theta oscillations by a working memory task. J. Neurosci. 21, 3175–3183 (2001).
Kempter, R., Gerstner, W. & Van Hemmen, J. L. Hebbian learning and spiking neurons. Phys. Rev. E 59, 4498 (1999).
Acknowledgements
This work was supported in part by the Israeli Science Foundation (grant number 346/16, O.B.; and grant number 155/18, N.B.). We thank Noam Ziv and Lukas Geyrhofer for helpful comments on an earlier version of this manuscript.
Author information
Authors and Affiliations
Contributions
All authors conceived the model. L.S. performed all the simulations and computations. O.B. and N.B. supervised the project. All authors wrote the manuscript.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing interests.
Additional information
Peer review information Nature Communications thanks Yonatan Loewenstein and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary information
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Susman, L., Brenner, N. & Barak, O. Stable memory with unstable synapses. Nat Commun 10, 4441 (2019). https://doi.org/10.1038/s41467019123062
Received:
Accepted:
Published:
DOI: https://doi.org/10.1038/s41467019123062
This article is cited by

STDPbased associative memory formation and retrieval
Journal of Mathematical Biology (2023)

An STDPbased encoding method for associative and composite data
Scientific Reports (2022)

Small, correlated changes in synaptic connectivity may facilitate rapid motor learning
Nature Communications (2022)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.